Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS20050114124 A1
Publication typeApplication
Application numberUS 10/724,008
Publication dateMay 26, 2005
Filing dateNov 26, 2003
Priority dateNov 26, 2003
Also published asCA2485800A1, CA2485800C, CA2786803A1, CN1622200A, CN1622200B, CN101887728A, CN101887728B, EP1536414A2, EP1536414A3, EP1536414B1, EP2431972A1, EP2431972B1, US7447630
Publication number10724008, 724008, US 2005/0114124 A1, US 2005/114124 A1, US 20050114124 A1, US 20050114124A1, US 2005114124 A1, US 2005114124A1, US-A1-20050114124, US-A1-2005114124, US2005/0114124A1, US2005/114124A1, US20050114124 A1, US20050114124A1, US2005114124 A1, US2005114124A1
InventorsZicheng Liu, Michael Sinclair, Alejandro Acero, Xuedong Huang, James Droppo, Li Deng, Zhengyou Zhang, Yanli Zheng
Original AssigneeMicrosoft Corporation
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
Method and apparatus for multi-sensory speech enhancement
US 20050114124 A1
Abstract
A method and system use an alternative sensor signal received from a sensor other than an air conduction microphone to estimate a clean speech value. The estimation uses either the alternative sensor signal alone, or in conjunction with the air conduction microphone signal. The clean speech value is estimated without using a model trained from noisy training data collected from an air conduction microphone. Under one embodiment, correction vectors are added to a vector formed from the alternative sensor signal in order to form a filter, which is applied to the air conductive microphone signal to produce the clean speech estimate. In other embodiments, the pitch of a speech signal is determined from the alternative sensor signal and is used to decompose an air conduction microphone signal. The decomposed signal is then used to determine a clean signal estimate.
Images(12)
Previous page
Next page
Claims(29)
1. A method of determining an estimate for a noise-reduced value representing a portion of a noise-reduced speech signal, the method comprising:
generating an alternative sensor signal using an alternative sensor other than an air conduction microphone;
converting the alternative sensor signal into at least one alternative sensor vector; and
adding a correction vector to the alternative sensor vector to form the estimate for the noise-reduced value.
2. The method of claim 1 wherein generating an alternative sensor signal comprises using a bone conduction microphone to generate the alternative sensor signal.
3. The method of claim 1 wherein adding a correction vector comprises adding a weighted sum of a plurality of correction vectors.
4. The method of claim 3 wherein each correction vector corresponds to a mixture component and each weight applied to a correction vector is based on the probability of the correction vector's mixture component given the alternative sensor vector.
5. The method of claim 1 further comprising training a correction vector through steps comprising:
generating an alternative sensor training signal;
converting the alternative sensor training signal into an alternative sensor training vector;
generating a clean air conduction microphone training signal;
converting the clean air conduction microphone training signal into an air conduction training vector; and
using the difference between the alternative sensor training vector and the air conduction training vector to form the correction vector.
6. The method of claim 5 wherein training a correction vector further comprises training a separate correction vector for each of a plurality of mixture components.
7. The method of claim 1 further comprising generating a refined estimate of a noise-reduced value through steps comprising:
generating an air conduction microphone signal;
converting the air conduction microphone signal into an air conduction vector;
estimating a noise value;
subtracting the noise value from the air conduction vector to form an air conduction estimate;
combining the air conduction estimate and the estimate for the noise-reduced value to form the refined estimate for the noise-reduced value.
8. The method of claim 7 wherein combining the air conduction estimate and the estimate for the noise-reduced value comprises combining the air conduction estimate and the estimate for the noise-reduced value in the power spectrum domain.
9. The method of claim 8 further comprising using the refined estimate for the noise-reduced value to form a filter.
10. The method of claim 1 wherein forming the estimate for the noise-reduced value comprises forming the estimate without estimating noise.
11. The method of claim 1 further comprising:
generating a second alternative sensor signal using a second alternative sensor other than an air conduction microphone;
converting the second alternative sensor signal into at least one second alternative sensor vector;
adding a correction vector to the second alternative sensor vector to form a second estimate for the noise-reduced value; and
combining the estimate for the noise-reduced value with the second estimate for the noise-reduced value to form a refined estimate for the noise-reduced value.
12. A method of determining an estimate of a clean speech value, the method comprising:
receiving an alternative sensor signal from a sensor other than an air conduction microphone;
receiving an air conduction microphone signal from an air conduction microphone;
identifying a pitch for a speech signal based on the alternative sensor signal;
using the pitch to decompose the air conduction microphone signal into a harmonic component and a residual component; and
using the harmonic component and the residual component to estimate the clean speech value.
13. The method of claim 12 wherein receiving an alternative sensor signal comprises receiving an alternative sensor signal from a bone conduction microphone.
14. A computer-readable medium having computer-executable instructions for performing steps comprising:
receiving an alternative sensor signal from an alternative sensor that is not an air conduction microphone; and
using the alternative sensor signal to estimate a clean speech value without using a model trained from noisy training data collected from an air conduction microphone.
15. The computer-readable medium of claim 14 wherein receiving an alternative sensor signal comprises receiving a sensor signal from a bone conduction microphone.
16. The computer-readable medium of claim 14 wherein using the alternative sensor signal to estimate a clean speech value comprises:
converting the alternative sensor signal into at least one alternative sensor vector; and
adding a correction vector to an alternative sensor vector.
17. The computer-readable medium of claim 16 wherein adding a correction vector comprises adding a weighted sum of a plurality of correction vectors, each correction vector being associated with a separate mixture component.
18. The computer-readable medium of claim 17 wherein adding a weighted sum of a plurality of correction vectors comprises using a weight that is based on the probability of a mixture component given the alternative sensor vector.
19. The computer-readable medium of claim 14 further comprising receiving a noisy test signal from an air conductive microphone and using the noisy test signal with the alternative sensor signal to estimate the clean speech value.
20. The computer-readable medium of claim 19 wherein using the noisy test signal comprises generating a noise model from the noisy test signal.
21. The computer-readable medium of claim 20 wherein using the noisy test signal further comprises:
converting the noisy test signal into at least one noisy test vector;
subtracting a mean of the noise model from the noisy test vector to form a difference; and
using the difference to estimate the clean speech value.
22. The computer-readable medium of claim 21 further comprising:
forming an alternative sensor vector from the alternative sensor signal;
adding a correction vector to the alternative sensor vector to form an alternative sensor estimate of the clean speech value; and
determining a weighted sum of the difference and the alternative sensor estimate to form the estimate of the clean speech value.
23. The computer-readable medium of claim 22 wherein the estimate of the clean speech value is in the power spectrum domain.
24. The computer-readable medium of claim 23 further comprising using the estimate of the clean speech value to form a filter.
25. The computer-readable medium of claim 14 wherein using the alternative sensor signal to estimate a clean speech value further comprises:
determining a pitch for a speech signal based on the alternative sensor signal; and
using the pitch to estimate the clean speech value.
26. The computer-readable medium of claim 25 wherein using the pitch to estimate the clean speech value comprises:
receiving a noisy test signal from an air conduction microphone; and
decomposing the noisy test signal into a harmonic component and a residual component based on the pitch.
27. The computer-readable medium of claim 26 further comprising using the harmonic component and the residual component to estimate the clean speech value.
28. The computer-readable medium of claim 14 wherein estimating a clean speech value further comprises not estimating noise.
29. The computer-readable medium of claim 14 further comprising:
receiving a second alternative sensor signal from a second alternative sensor that is not an air conduction microphone; and
using the second alternative sensor signal with the alternative sensor signal to estimate the clean speech value.
Description
BACKGROUND OF THE INVENTION

The present invention relates to noise reduction. In particular, the present invention relates to removing noise from speech signals.

A common problem in speech recognition and speech transmission is the corruption of the speech signal by additive noise. In particular, corruption due to the speech of another speaker has proven to be difficult to detect and/or correct.

One technique for removing noise attempts to model the noise using a set of noisy training signals collected under various conditions. These training signals are received before a test signal that is to be decoded or transmitted and are used for training purposes only. Although such systems attempt to build models that take noise into consideration, they are only effective if the noise conditions of the training signals match the noise conditions of the test signals. Because of the large number of possible noises and the seemingly infinite combinations of noises, it is very difficult to build noise models from training signals that can handle every test condition.

Another technique for removing noise is to estimate the noise in the test signal and then subtract it from the noisy speech signal. Typically, such systems estimate the noise from previous frames of the test signal. As such, if the noise is changing over time, the estimate of the noise for the current frame will be inaccurate.

One system of the prior art for estimating the noise in a speech signal uses the harmonics of human speech. The harmonics of human speech produce peaks in the frequency spectrum. By identifying nulls between these peaks, these systems identify the spectrum of the noise. This spectrum is then subtracted from the spectrum of the noisy speech signal to provide a clean speech signal.

The harmonics of speech have also been used in speech coding to reduce the amount of data that must be sent when encoding speech for transmission across a digital communication path. Such systems attempt to separate the speech signal into a harmonic component and a random component. Each component is then encoded separately for transmission. One system in particular used a harmonic+noise model in which a sum-of-sinusoids model is fit to the speech signal to perform the decomposition.

In speech coding, the decomposition is done to find a parameterization of the speech signal that accurately represents the input noisy speech signal. The decomposition has no noise-reduction capability.

Recently, a system has been developed that attempts to remove noise by using a combination of an alternative sensor, such as a bone conduction microphone, and an air conduction microphone. This system is trained using three training channels: a noisy alternative sensor training signal, a noisy air conduction microphone training signal, and a clean air conduction microphone training signal. Each of the signals is converted into a feature domain. The features for the noisy alternative sensor signal and the noisy air conduction microphone signal are combined into a single vector representing a noisy signal. The features for the clean air conduction microphone signal form a single clean vector. These vectors are then used to train a mapping between the noisy vectors and the clean vectors. Once trained, the mappings are applied to a noisy vector formed from a combination of a noisy alternative sensor test signal and a noisy air conduction microphone test signal. This mapping produces a clean signal vector.

This system is less than optimum when the noise conditions of the test signals do not match the noise conditions of the training signals because the mappings are designed for the noise conditions of the training signals.

SUMMARY OF THE INVENTION

A method and system use an alternative sensor signal received from a sensor other than an air conduction microphone to estimate a clean speech value. The clean speech value is estimated without using a model trained from noisy training data collected from an air conduction microphone. Under one embodiment, correction vectors are added to a vector formed from the alternative sensor signal in order to form a filter, which is applied to the air conductive microphone signal to produce the clean speech estimate. In other embodiments, the pitch of a speech signal is determined from the alternative sensor signal and is used to decompose an air conduction microphone signal. The decomposed signal is then used to identify a clean signal estimate.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of one computing environment in which the present invention may be practiced.

FIG. 2 is a block diagram of an alternative computing environment in which the present invention may be practiced.

FIG. 3 is a block diagram of a general speech processing system of the present invention.

FIG. 4 is a block diagram of a system for training noise reduction parameters under one embodiment of the present invention.

FIG. 5 is a flow diagram for training noise reduction parameters using the system of FIG. 4.

FIG. 6 is a block diagram of a system for identifying an estimate of a clean speech signal from a noisy test speech signal under one embodiment of the present invention.

FIG. 7 is a flow diagram of a method for identifying an estimate of a clean speech signal using the system of FIG. 6.

FIG. 8 is a block diagram of an alternative system for identifying an estimate of a clean speech signal.

FIG. 9 is a block diagram of a second alternative system for identifying an estimate of a clean speech signal.

FIG. 10 is a flow diagram of a method for identifying an estimate of a clean speech signal using the system of FIG. 9.

FIG. 11 is a block diagram of a bone conduction microphone.

DETAILED DESCRIPTION OF ILLUSTRATIVE EMBODIMENTS

FIG. 1 illustrates an example of a suitable computing system environment 100 on which the invention may be implemented. The computing system environment 100 is only one example of a suitable computing environment and is not intended to suggest any limitation as to the scope of use or functionality of the invention. Neither should the computing environment 100 be interpreted as having any dependency or requirement relating to any one or combination of components illustrated in the exemplary operating environment 100.

The invention is operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well-known computing systems, environments, and/or configurations that may be suitable for use with the invention include, but are not limited to, personal computers, server computers, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, telephony systems, distributed computing environments that include any of the above systems or devices, and the like.

The invention may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The invention is designed to be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules are located in both local and remote computer storage media including memory storage devices.

With reference to FIG. 1, an exemplary system for implementing the invention includes a general-purpose computing device in the form of a computer 110. Components of computer 110 may include, but are not limited to, a processing unit 120, a system memory 130, and a system bus 121 that couples various system components including the system memory to the processing unit 120. The system bus 121 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. By way of example, and not limitation, such architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus also known as Mezzanine bus.

Computer 110 typically includes a variety of computer readable media. Computer readable media can be any available media that can be accessed by computer 110 and includes both volatile and nonvolatile media, removable and non-removable media. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by computer 110. Communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared and other wireless media. Combinations of any of the above should also be included within the scope of computer readable media.

The system memory 130 includes computer storage media in the form of volatile and/or nonvolatile memory such as read only memory (ROM) 131 and random access memory (RAM) 132. A basic input/output system 133 (BIOS), containing the basic routines that help to transfer information between elements within computer 110, such as during start-up, is typically stored in ROM 131. RAM 132 typically contains data and/or program modules that are immediately accessible to and/or presently being operated on by processing unit 120. By way of example, and not limitation, FIG. 1 illustrates operating system 134, application programs 135, other program modules 136, and program data 137.

The computer 110 may also include other removable/non-removable volatile/nonvolatile computer storage media. By way of example only, FIG. 1 illustrates a hard disk drive 141 that reads from or writes to non-removable, nonvolatile magnetic media, a magnetic disk drive 151 that reads from or writes to a removable, nonvolatile magnetic disk 152, and an optical disk drive 155 that reads from or writes to a removable, nonvolatile optical disk 156 such as a CD ROM or other optical media. Other removable/non-removable, volatile/nonvolatile computer storage media that can be used in the exemplary operating environment include, but are not limited to, magnetic tape cassettes, flash memory cards, digital versatile disks, digital video tape, solid state RAM, solid state ROM, and the like. The hard disk drive 141 is typically connected to the system bus 121 through a non-removable memory interface such as interface 140, and magnetic disk drive 151 and optical disk drive 155 are typically connected to the system bus 121 by a removable memory interface, such as interface 150.

The drives and their associated computer storage media discussed above and illustrated in FIG. 1, provide storage of computer readable instructions, data structures, program modules and other data for the computer 110. In FIG. 1, for example, hard disk drive 141 is illustrated as storing operating system 144, application programs 145, other program modules 146, and program data 147. Note that these components can either be the same as or different from operating system 134, application programs 135, other program modules 136, and program data 137. Operating system 144, application programs 145, other program modules 146, and program data 147 are given different numbers here to illustrate that, at a minimum, they are different copies.

A user may enter commands and information into the computer 110 through input devices such as a keyboard 162, a microphone 163, and a pointing device 161, such as a mouse, trackball or touch pad. Other input devices (not shown) may include a joystick, game pad, satellite dish, scanner, or the like. These and other input devices are often connected to the processing unit 120 through a user input interface 160 that is coupled to the system bus, but may be connected by other interface and bus structures, such as a parallel port, game port or a universal serial bus (USB). A monitor 191 or other type of display device is also connected to the system bus 121 via an interface, such as a video interface 190. In addition to the monitor, computers may also include other peripheral output devices such as speakers 197 and printer 196, which may be connected through an output peripheral interface 195.

The computer 110 is operated in a networked environment using logical connections to one or more remote computers, such as a remote computer 180. The remote computer 180 may be a personal computer, a hand-held device, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to the computer 110. The logical connections depicted in FIG. 1 include a local area network (LAN) 171 and a wide area network (WAN) 173, but may also include other networks. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.

When used in a LAN networking environment, the computer 110 is connected to the LAN 171 through a network interface or adapter 170. When used in a WAN networking environment, the computer 110 typically includes a modem 172 or other means for establishing communications over the WAN 173, such as the Internet. The modem 172, which may be internal or external, may be connected to the system bus 121 via the user input interface 160, or other appropriate mechanism. In a networked environment, program modules depicted relative to the computer 110, or portions thereof, may be stored in the remote memory storage device. By way of example, and not limitation, FIG. 1 illustrates remote application programs 185 as residing on remote computer 180. It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between the computers may be used.

FIG. 2 is a block diagram of a mobile device 200, which is an exemplary computing environment. Mobile device 200 includes a microprocessor 202, memory 204, input/output (I/O) components 206, and a communication interface 208 for communicating with remote computers or other mobile devices. In one embodiment, the afore-mentioned components are coupled for communication with one another over a suitable bus 210.

Memory 204 is implemented as non-volatile electronic memory such as random access memory (RAM) with a battery back-up module (not shown) such that information stored in memory 204 is not lost when the general power to mobile device 200 is shut down. A portion of memory 204 is preferably allocated as addressable memory for program execution, while another portion of memory 204 is preferably used for storage, such as to simulate storage on a disk drive.

Memory 204 includes an operating system 212, application programs 214 as well as an object store 216. During operation, operating system 212 is preferably executed by processor 202 from memory 204. Operating system 212, in one preferred embodiment, is a WINDOWS® CE brand operating system commercially available from Microsoft Corporation. Operating system 212 is preferably designed for mobile devices, and implements database features that can be utilized by applications 214 through a set of exposed application programming interfaces and methods. The objects in object store 216 are maintained by applications 214 and operating system 212, at least partially in response to calls to the exposed application programming interfaces and methods.

Communication interface 208 represents numerous devices and technologies that allow mobile device 200 to send and receive information. The devices include wired and wireless modems, satellite receivers and broadcast tuners to name a few. Mobile device 200 can also be directly connected to a computer to exchange data therewith. In such cases, communication interface 208 can be an infrared transceiver or a serial or parallel communication connection, all of which are capable of transmitting streaming information.

Input/output components 206 include a variety of input devices such as a touch-sensitive screen, buttons, rollers, and a microphone as well as a variety of output devices including an audio generator, a vibrating device, and a display. The devices listed above are by way of example and need not all be present on mobile device 200. In addition, other input/output devices may be attached to or found with mobile device 200 within the scope of the present invention.

FIG. 3 provides a basic block diagram of embodiments of the present invention. In FIG. 3, a speaker 300 generates a speech signal 302 that is detected by an air conduction microphone 304 and an alternative sensor 306. Examples of alternative sensors include a throat microphone that measures the user's throat vibrations, a bone conduction sensor that is located on or adjacent to a facial or skull bone of the user (such as the jaw bone) or in the ear of the user and that senses vibrations of the skull and jaw that correspond to speech generated by the user. Air conduction microphone 304 is the type of microphone that is used commonly to convert audio air-waves into electrical signals.

Air conduction microphone 304 also receives noise 308 generated by one or more noise sources 310. Depending on the type of alternative sensor and the level of the noise, noise 308 may also be detected by alternative sensor 306. However, under embodiments of the present invention, alternative sensor 306 is typically less sensitive to ambient noise than air conduction microphone 304. Thus, the alternative sensor signal 312 generated by alternative sensor 306 generally includes less noise than air conduction microphone signal 314 generated by air conduction microphone 304.

Alternative sensor signal 312 and air conduction microphone signal 314 are provided to a clean signal estimator 316, which estimates a clean signal 318. Clean signal estimate 318 is provided to a speech process 320. Clean signal estimate 318 may either be a filtered time-domain signal or a feature domain vector. If clean signal estimate 318 is a time-domain signal, speech process 320 may take the form of a listener, a speech coding system, or a speech recognition system. If clean signal estimate 318 is a feature domain vector, speech process 320 will typically be a speech recognition system.

The present invention provides several methods and systems for estimating clean speech using air conduction microphone signal 314 and alternative sensor signal 312. One system uses stereo training data to train correction vectors for the alternative sensor signal. When these correction vectors are later added to a test alternative sensor vector, they provide an estimate of a clean signal vector. One further extension of this system is to first track time-varying distortion and then to incorporate this information into the computation of the correction vectors and into the estimation of clean speech.

A second system provides an interpolation between the clean signal estimate generated by the correction vectors and an estimate formed by subtracting an estimate of the current noise in the air conduction test signal from the air conduction signal. A third system uses the alternative sensor signal to estimate the pitch of the speech signal and then uses the estimated pitch to identify an estimate for the clean signal. Each of these systems is discussed separately below.

Training Stereo Correction Vectors

FIGS. 4 and 5 provide a block diagram and flow diagram for training stereo correction vectors for the two embodiments of the present invention that rely on correction vectors to generate an estimate of clean speech.

The method of identifying correction vectors begins in step 500 of FIG. 5, where a “clean” air conduction microphone signal is converted into a sequence of feature vectors. To do this, a speaker 400 of FIG. 4, speaks into an air conduction microphone 410, which converts the audio waves into electrical signals. The electrical signals are then sampled by an analog-to-digital converter 414 to generate a sequence of digital values, which are grouped into frames of values by a frame constructor 416. In one embodiment, A-to-D converter 414 samples the analog signal at 16 kHz and 16 bits per sample, thereby creating 32 kilobytes of speech data per second and frame constructor 416 creates a new frame every 10 milliseconds that includes 25 milliseconds worth of data.

Each frame of data provided by frame constructor 416 is converted into a feature vector by a feature extractor 418. Under one embodiment, feature extractor 418 forms cepstral features. Examples of such features include LPC derived cepstrum, and Mel-Frequency Cepstrum Coefficients. Examples of other possible feature extraction modules that may be used with the present invention include modules for performing Linear Predictive Coding (LPC), Perceptive Linear Prediction (PLP), and Auditory model feature extraction. Note that the invention is not limited to these feature extraction modules and that other modules may be used within the context of the present invention.

In step 502 of FIG. 5, an alternative sensor signal is converted into feature vectors. Although the conversion of step 502 is shown as occurring after the conversion of step 500, any part of the conversion may be performed before, during or after step 500 under the present invention. The conversion of step 502 is performed through a process similar to that described above for step 500.

In the embodiment of FIG. 4, this process begins when alternative sensor 402 detects a physical event associated with the production of speech by speaker 400 such as bone vibration or facial movement. As shown in FIG. 11, in one embodiment of a bone conduction sensor 1100, a soft elastomer bridge 1102 is adhered to the diaphragm 1104 of a normal air conduction microphone 1106. This soft bridge 1102 conducts vibrations from skin contact 1108 of the user directly to the diaphragm 1104 of microphone 1106. The movement of diaphragm 1104 is converted into an electrical signal by a transducer 1110 in microphone 1106. Alternative sensor 402 converts the physical event into analog electrical signal, which is sampled by an analog-to-digital converter 404. The sampling characteristics for A/D converter 404 are the same as those described above for A/D converter 414. The samples provided by A/D converter 404 are collected into frames by a frame constructor 406, which acts in a manner similar to frame constructor 416. These frames of samples are then converted into feature vectors by a feature extractor 408, which uses the same feature extraction method as feature extractor 418.

The feature vectors for the alternative sensor signal and the air conductive signal are provided to a noise reduction trainer 420 in FIG. 4. At step 504 of FIG. 5, noise reduction trainer 420 groups the feature vectors for the alternative sensor signal into mixture components. This grouping can be done by grouping similar feature vectors together using a maximum likelihood training technique or by grouping feature vectors that represent a temporal section of the speech signal together. Those skilled in the art will recognize that other techniques for grouping the feature vectors may be used and that the two techniques listed above are only provided as examples.

Noise reduction trainer 420 then determines a correction vector, rs, for each mixture component, s, at step 508 of FIG. 5. Under one embodiment, the correction vector for each mixture component is determined using maximum likelihood criterion. Under this technique, the correction vector is calculated as: r s = t p ( s b t ) ( x t - b t ) t p ( s b t ) EQ . 1

Where xt is the value of the air conduction vector for frame t and bt is the value of the alternative sensor vector for frame t. In Equation 1: p ( s b t ) = p ( b t s ) p ( s ) s p ( b t s ) p ( s ) EQ . 2
where p(s) is simply one over the number of mixture components and p(bt|s) is modeled as a Gaussian distribution:
p(b t |s)=N(b tbb)  EQ. 3
with the mean μb and variance Γb trained using an Expectation Maximization (EM) algorithm where each iteration consists of the following steps: γ s ( t ) = p ( s b t ) EQ . 4 μ s = t γ s ( t ) b t t γ s ( t ) EQ . 5 Γ s = t γ s ( t ) ( b t - μ s ) ( b t - μ s ) T t γ s ( t ) EQ . 6
EQ. 4 is the E-step in the EM algorithm, which uses the previously estimated parameters. EQ. 5 and EQ. 6 are the M-step, which updates the parameters using the E-step results.

The E- and M-steps of the algorithm iterate until stable values for the model parameters are determined. These parameters are then used to evaluate equation 1 to form the correction vectors. The correction vectors and the model parameters are then stored in a noise reduction parameter storage 422.

After a correction vector has been determined for each mixture component at step 508, the process of training the noise reduction system of the present invention is complete. Once a correction vector has been determined for each mixture, the vectors may be used in a noise reduction technique of the present invention. Two separate noise reduction techniques that use the correction vectors are discussed below.

Noise Reduction using Correction Vector and Noise Estimate

A system and method that reduces noise in a noisy speech signal based on correction vectors and a noise estimate is shown in the block diagram of FIG. 6 and the flow diagram of FIG. 7, respectively.

At step 700, an audio test signal detected by an air conduction microphone 604 is converted into feature vectors. The audio test signal received by microphone 604 includes speech from a speaker 600 and additive noise from one or more noise sources 602. The audio test signal detected by microphone 604 is converted into an electrical signal that is provided to analog-to-digital converter 606.

A-to-D converter 606 converts the analog signal from microphone 604 into a series of digital values. In several embodiments, A-to-D converter 606 samples the analog signal at 16 kHz and 16 bits per sample, thereby creating 32 kilobytes of speech data per second. These digital values are provided to a frame constructor 607, which, in one embodiment, groups the values into 25 millisecond frames that start 10 milliseconds apart.

The frames of data created by frame constructor 607 are provided to feature extractor 610, which extracts a feature from each frame. Under one embodiment, this feature extractor is different from feature extractors 408 and 418 that were used to train the correction vectors. In particular, under this embodiment, feature extractor 610 produces power spectrum values instead of cepstral values. The extracted features are provided to a clean signal estimator 622, a speech detection unit 626 and a noise model trainer 624.

At step 702, a physical event, such as bone vibration or facial movement, associated with the production of speech by speaker 600 is converted into a feature vector. Although shown as a separate step in FIG. 7, those skilled in the art will recognize that portions of this step may be done at the same time as step 700. During step 702, the physical event is detected by alternative sensor 614. Alternative sensor 614 generates an analog electrical signal based on the physical events. This analog signal is converted into a digital signal by analog-to-digital converter 616 and the resulting digital samples are grouped into frames by frame constructor 617. Under one embodiment, analog-to-digital converter 616 and frame constructor 617 operate in a manner similar to analog-to-digital converter 606 and frame constructor 607.

The frames of digital values are provided to a feature extractor 620, which uses the same feature extraction technique that was used to train the correction vectors. As mentioned above, examples of such feature extraction modules include modules for performing Linear Predictive Coding (LPC), LPC derived cepstrum, Perceptive Linear Prediction (PLP), Auditory model feature extraction, and Mel-Frequency Cepstrum Coefficients (MFCC) feature extraction. In many embodiments, however, feature extraction techniques that produce cepstral features are used.

The feature extraction module produces a stream of feature vectors that are each associated with a separate frame of the speech signal. This stream of feature vectors is provided to clean signal estimator 622.

The frames of values from frame constructor 617 are also provided to a feature extractor 621, which in one embodiment extracts the energy of each frame. The energy value for each frame is provided to a speech detection unit 626.

At step 704, speech detection unit 626 uses the energy feature of the alternative sensor signal to determine when speech is likely present. This information is passed to noise model trainer 624, which attempts to model the noise during periods when there is no speech at step 706.

Under one embodiment, speech detection unit 626 first searches the sequence of frame energy values to find a peak in the energy. It then searches for a valley after the peak. The energy of this valley is referred to as an energy separator, d. To determine if a frame contains speech, the ratio, k, of the energy of the frame, e, over the energy separator, d, is then determined as: k=e/d. A speech confidence, q, for the frame is then determined as: q = { 0 : k < 1 k - 1 α - 1 : 1 k α 1 : k > α EQ . 7
where α defines the transition between two states and in one implementation is set to 2. Finally, we use the average confidence value of its 5 neighboring frames (including itself) as the final confidence value for this frame.

Under one embodiment, a fixed threshold value is used to determine if speech is present such that if the confidence value exceeds the threshold, the frame is considered to contain speech and if the confidence value does not exceed the threshold, the frame is considered to contain non-speech. Under one embodiment, a threshold value of 0.1 is used.

For each non-speech frame detected by speech detection unit 626, noise model trainer 624 updates a noise model 625 at step 706. Under one embodiment, noise model 625 is a Gaussian model that has a mean μn and a variance Σn. This model is based on a moving window of the most recent frames of non-speech. Techniques for determining the mean and variance from the non-speech frames in the window are well known in the art.

Correction vectors and model parameters in parameter storage 422 and noise model 625 are provided to clean signal estimator 622 with the feature vectors, b, for the alternative sensor and the feature vectors, Sy, for the noisy air conduction microphone signal. At step 708, clean signal estimator 622 estimates an initial value for the clean speech signal based on the alternative sensor feature vector, the correction vectors, and the model parameters for the alternative sensor. In particular, the alternative sensor estimate of the clean signal is calculated as: x ^ = b + s p ( s | b ) r s EQ . 8
where {circumflex over (x)} is the clean signal estimate in the cepstral domain, b is the alternative sensor feature vector, p(s|b) is determined using equation 2 above, and rs is the correction vector for mixture component s. Thus, the estimate of the clean signal in Equation 8 is formed by adding the alternative sensor feature vector to a weighted sum of correction vectors where the weights are based on the probability of a mixture component given the alternative sensor feature vector.

At step 710, the initial alternative sensor clean speech estimate is refined by combining it with a clean speech estimate that is formed from the noisy air conduction microphone vector and the noise model. This results in a refined clean speech estimate 628. In order to combine the cepstral value of the initial clean signal estimate with the power spectrum feature vector of the noisy air conduction microphone, the cepstral value is converted to the power spectrum domain using:
Ŝ x|b =e C −1 {circumflex over (x)}   EQ. 9
where C−1 is an inverse discrete cosine transform and Ŝx|b is the power spectrum estimate of the clean signal based on the alternative sensor.

Once the initial clean signal estimate from the alternative sensor has been placed in the power spectrum domain, it can be combined with the noisy air conduction microphone vector and the noise model as:
Ŝ x=(Σn −1x|b −1)−1n −1(S y−μn)+Σx|b −1 Ŝ x|b]  EQ. 10
where Ŝx is the refined clean signal estimate in the power spectrum domain, Sy is the noisy air conduction microphone feature vector, (μnn) are the mean and covariance of the prior noise model (see 624), Ŝx|b is the initial clean signal estimate based on the alternative sensor, and Σx|b is the covariance matrix of the conditional probability distribution for the clean speech given the alternative sensor's measurement. Σx|b can be computed as follows. Let J denote the Jacobian of the function on the right hand side of equation 9. Let Σ be the covariance matrix of {circumflex over (x)}. Then the covariance of Ŝx|b is
Σx|b =JΣJ T  EQ. 11

In a simplified embodiment, we rewrite EQ. 10 as the following equation:
Ŝ x=α(f)(S y−μn)+(1−α(f))Ŝ x|b  EQ. 12
where α(f) is a function of both the time and the frequency band. Since the alternative sensor that we are currently using has the bandwidth up to 3 KHz, we choose α(f) to be 0 for the frequency band below 3 KHz. Basically, we trust the initial clean signal estimate from the alternative sensor for low frequency bands. For high frequency bands, the initial clean signal estimate from the alterative sensor is not so reliable. Intuitively, when the noise is small for a frequency band at the current frame, we would like to choose a large α(f) so that we use more information from the air conduction microphone for this frequency band. Otherwise, we would like to use more information from the alternative sensor by choosing a small α(f). In one embodiment, we use the energy of the initial clean signal estimate from the alternative sensor to determine the noise level for each frequency band. Let E(f) denote the energy for frequency band f. Let M=MaxfE(f). α(f), as a function of ft is defined as follows: α ( f ) = { E ( f ) M : f 4 K f - 3 K 1 K α ( 4 K ) : 3 K < f < 4 K 0 : f 3 K EQ . 13
where we use a linear interpolation to transition from 3K to 4K to ensure the smoothness of α(f).

The refined clean signal estimate in the power spectrum domain may be used to construct a Wiener filter to filter the noisy air conduction microphone signal. In particular, the Wiener filter, H, is set such that: H = S ^ x S y EQ . 14

This filter can then be applied against the time domain noisy air conduction microphone signal to produce a noise-reduced or clean time-domain signal. The noise-reduced signal can be provided to a listener or applied to a speech recognizer.

Note that Equation 12 provides a refined clean signal estimate that is the weighted sum of two factors, one of which is a clean signal estimate from an alternative sensor. This weighted sum can be extended to include additional factors for additional alternative sensors. Thus, more than one alternate sensor may be used to generate independent estimates of the clean signal. These multiple estimates can then be combined using equation 12.

Noise Reduction using Correction Vector without Noise Estimate

FIG. 8 provides a block diagram of an alternative system for estimating a clean speech value under the present invention. The system of FIG. 8 is similar to the system of FIG. 6 except that the estimate of the clean speech value is formed without the need for an air conduction microphone or a noise model.

In FIG. 8, a physical event associated with a speaker 800 producing speech is converted into a feature vector by alternative sensor 802, analog-to-digital converter 804, frame constructor 806 and feature extractor 808, in a manner similar to that discussed above for alternative sensor 614, analog-to-digital converter 616, frame constructor 617 and feature extractor 618 of FIG. 6. The feature vectors from feature extractor 808 and the noise reduction parameters 422 are provided to a clean signal estimator 810, which determines an estimate of a clean signal value 812, Ŝx|b, using equations 8 and 9 above.

The clean signal estimate, Ŝx|b, in the power spectrum domain may be used to construct a Wiener filter to filter a noisy air conduction microphone signal. In particular, the Wiener filter, H, is set such that: H = S ^ x b S y EQ . 15

This filter can then be applied against the time domain noisy air conduction microphone signal to produce a noise-reduced or clean signal. The noise-reduced signal can be provided to a listener or applied to a speech recognizer.

Alternatively, the clean signal estimate in the cepstral domain, {circumflex over (x)}, which is calculated in Equation 8, may be applied directly to a speech recognition system.

Noise Reduction using Pitch Tracking

An alternative technique for generating estimates of a clean speech signal is shown in the block diagram of FIG. 9 and the flow diagram of FIG. 10. In particular, the embodiment of FIGS. 9 and 10 determine a clean speech estimate by identifying a pitch for the speech signal using an alternative sensor and then using the pitch to decompose a noisy air conduction microphone signal into a harmonic component and a random component. Thus, the noisy signal is represented as:
y=y h +y r  EQ. 16
where y is the noisy signal, yh is the harmonic component, and yr is the random component. A weighted sum of the harmonic component and the random component are used to form a noise-reduced feature vector representing a noise-reduced speech signal.

Under one embodiment, the harmonic component is modeled as a sum of harmonically-related sinusoids such that: y h = k = 1 K a k cos ( k ω 0 t ) + b k sin ( k ω 0 t ) EQ . 17
where Ω0 is the fundamental or pitch frequency and K is the total number of harmonics in the signal.

Thus, to identify the harmonic component, an estimate of the pitch frequency and the amplitude parameters {a1a2 . . . akb1b2 . . . bk} must be determined.

At step 1000, a noisy speech signal is collected and converted into digital samples. To do this, an air conduction microphone 904 converts audio waves from a speaker 900 and one or more additive noise sources 902 into electrical signals. The electrical signals are then sampled by an analog-to-digital converter 906 to generate a sequence of digital values. In one embodiment, A-to-D converter 906 samples the analog signal at 16 kHz and 16 bits per sample, thereby creating 32 kilobytes of speech data per second. At step 1002, the digital samples are grouped into frames by a frame constructor 908. Under one embodiment, frame constructor 908 creates a new frame every 10 milliseconds that includes 25 milliseconds worth of data.

At step 1004, a physical event associated with the production of speech is detected by alternative sensor 944. In this embodiment, an alternative sensor that is able to detect harmonic components, such as a bone conduction sensor, is best suited to be used as alternative sensor 944. Note that although step 1004 is shown as being separate from step 1000, those skilled in the art will recognize that these steps may be performed at the same time. The analog signal generated by alternative sensor 944 is converted into digital samples by an analog-to-digital converter 946. The digital samples are then grouped into frames by a frame constructer 948 at step 1006.

At step 1008, the frames of the alternative sensor signal are used by a pitch tracker 950 to identify the pitch or fundamental frequency of the speech.

An estimate for the pitch frequency can be determined using any number of available pitch tracking systems. Under many of these systems, candidate pitches are used to identify possible spacing between the centers of segments of the alternative sensor signal. For each candidate pitch, a correlation is determined between successive segments of speech. In general, the candidate pitch that provides the best correlation will be the pitch frequency of the frame. In some systems, additional information is used to refine the pitch selection such as the energy of the signal and/or an expected pitch track.

Given an estimate of the pitch from pitch tracker 950, the air conduction signal vector can be decomposed into a harmonic component and a random component at step 1010. To do so, equation 17 is rewritten as:
y=Ab  EQ. 18
where y is a vector of N samples of the noisy speech signal, A is an N×2K matrix given by:
A=[A cos A sin]  EQ. 19
with elements
A cos(k,t)=cos( 0 t) A sin(k,t)=sin( 0 t)  EQ. 20
and b is a 2K×1 vector given by:
b T =[a 1 a 2 . . . a k b 1 b 2 . . . b k]  EQ. 21
Then, the least-squares solution for the amplitude coefficients is:
{circumflex over (b)}=(A T A)−1 A T y  EQ. 22
Using {circumflex over (b)}, an estimate for the harmonic component of the noisy speech signal can be determined as:
y h =A{circumflex over (b)}  EQ. 23

An estimate of the random component is then calculated as:
y r =y−y h  EQ. 24

Thus, using equations 18-24 above, harmonic decompose unit 910 is able to produce a vector of harmonic component samples 912, yh, and a vector of random component samples 914, yr.

After the samples of the frame have been decomposed into harmonic and random samples, a scaling parameter or weight is determined for the harmonic component at step 1012. This scaling parameter is used as part of a calculation of a noise-reduced speech signal as discussed further below. Under one embodiment, the scaling parameter is calculated as: α h = i y h ( i ) 2 i y ( i ) 2 EQ . 25
where αh is the scaling parameter, yh(t) is the ith sample in the vector of harmonic component samples yh and y(i) is the ith sample of the noisy speech signal for this frame. In Equation 25, the numerator is the sum of the energy of each sample of the harmonic component and the denominator is the sum of the energy of each sample of the noisy speech signal. Thus, the scaling parameter is the ratio of the harmonic energy of the frame to the total energy of the frame.

In alternative embodiments, the scaling parameter is set using a probabilistic voiced-unvoiced detection unit. Such units provide the probability that a particular frame of speech is voiced, meaning that the vocal cords resonate during the frame, rather than unvoiced. The probability that the frame is from a voiced region of speech can be used directly as the scaling parameter.

After the scaling parameter has been determined or while it is being determined, the Mel spectra for the vector of harmonic component samples and the vector of random component samples are determined at step 1014. This involves passing each vector of samples through a Discrete Fourier Transform (DFT) 918 to produce a vector of harmonic component frequency values 922 and a vector of random component frequency values 920. The power spectra represented by the vectors of frequency values are then smoothed by a Mel weighting unit 924 using a series of triangular weighting functions applied along the Mel scale. This results in a harmonic component Mel spectral vector 928, Yh, and a random component Mel spectral vector 926, Yr.

At step 1016, the Mel spectra for the harmonic component and the random component are combined as a weighted sum to form an estimate of a noise-reduced Mel spectrum. This step is performed by weighted sum calculator 930 using the scaling factor determined above in the following equation:
{circumflex over (X)}(t)=αh(t)Y h(t)+αr Y r(t)  EQ. 26
where {circumflex over (X)}(t) is the estimate of the noise-reduced Mel spectrum, Yh(t) is the harmonic component Mel spectrum, Yr(t) is the random component Mel spectrum, αh(t) is the scaling factor determined above, αr is a fixed scaling factor for the random component that in one embodiment is set equal to 0.1, and the time index t is used to emphasize that the scaling factor for the harmonic component is determined for each frame while the scaling factor for the random component remains fixed. Note that in other embodiments, the scaling factor for the random component may be determined for each frame.

After the noise-reduced Mel spectrum has been calculated at step 1016, the log 932 of the Mel spectrum is determined and then is applied to a Discrete Cosine Transform 934 at step 1018. This produces a Mel Frequency Cepstral Coefficient (MFCC) feature vector 936 that represents a noise-reduced speech signal.

A separate noise-reduced MFCC feature vector is produced for each frame of the noisy signal. These feature vectors may be used for any desired purpose including speech enhancement and speech recognition. For speech enhancement, the MFCC feature vectors can be converted into the power spectrum domain and can be used with the noisy air conduction signal to form a Weiner filter.

Although the present invention has been described with reference to particular embodiments, workers skilled in the art will recognize that changes may be made in form and detail without departing from the spirit and scope of the invention.

Referenced by
Citing PatentFiling datePublication dateApplicantTitle
US7283850 *Oct 12, 2004Oct 16, 2007Microsoft CorporationMethod and apparatus for multi-sensory speech enhancement on a mobile device
US7346504Jun 20, 2005Mar 18, 2008Microsoft CorporationMulti-sensory speech enhancement using a clean speech prior
US7356471 *Jun 23, 2003Apr 8, 2008Denso CorporationAdjusting sound characteristic of a communication network using test signal prior to providing communication to speech recognition server
US7383181Jul 29, 2003Jun 3, 2008Microsoft CorporationMulti-sensory speech detection system
US7406303Sep 16, 2005Jul 29, 2008Microsoft CorporationMulti-sensory speech enhancement using synthesized sensor signal
US7499686Feb 24, 2004Mar 3, 2009Microsoft CorporationMethod and apparatus for multi-sensory speech enhancement on a mobile device
US7574008Sep 17, 2004Aug 11, 2009Microsoft CorporationMethod and apparatus for multi-sensory speech enhancement
US7680656Jun 28, 2005Mar 16, 2010Microsoft CorporationMulti-sensory speech enhancement using a speech-state model
US7930178Dec 23, 2005Apr 19, 2011Microsoft CorporationSpeech modeling and enhancement based on magnitude-normalized spectra
US8019089 *Nov 20, 2006Sep 13, 2011Microsoft CorporationRemoval of noise, corresponding to user input devices from an audio signal
US8370139Mar 19, 2007Feb 5, 2013Kabushiki Kaisha ToshibaFeature-vector compensating apparatus, feature-vector compensating method, and computer program product
US8645132 *Aug 24, 2011Feb 4, 2014Sensory, Inc.Truly handsfree speech recognition in high noise environments
US8660849 *Dec 21, 2012Feb 25, 2014Apple Inc.Prioritizing selection criteria by automated assistant
US20130054235 *Aug 24, 2011Feb 28, 2013Sensory, IncorporatedTruly handsfree speech recognition in high noise environments
Classifications
U.S. Classification704/228, 704/E21.004
International ClassificationG10L21/02
Cooperative ClassificationG10L2021/02165, G10L21/0208
European ClassificationG10L21/0208
Legal Events
DateCodeEventDescription
Apr 11, 2012FPAYFee payment
Year of fee payment: 4
Jul 7, 2004ASAssignment
Owner name: MICROSOFT CORPORATION, WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:LIU, ZICHENG;SINCLAIR, MICHAEL J.;ACERO, ALEJANDRO;AND OTHERS;REEL/FRAME:014824/0933;SIGNING DATES FROM 20031218 TO 20040121
Jul 2, 2004ASAssignment
Owner name: MICROSOFT CORPORATION, WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:LIN, ZICHENG;SINCLAIR, MICHAEL J.;ACERO, ALEJANDRO;AND OTHERS;REEL/FRAME:014814/0234;SIGNING DATES FROM 20031218 TO 20040121
Mar 8, 2004ASAssignment
Owner name: MICROSOFT CORPORATION, WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:LIU, ZICHENG;SINCLAIR, MICHAEL J.;ACERO, ALEJANDRO;AND OTHERS;REEL/FRAME:015046/0696;SIGNING DATES FROM 20031218 TO 20040121